Agüera y Arcas: life is self-constructing computation that grows complex through cooperation, and his team watched replicators appear in random code

A definition of life: self-constructing computation that complexifies through symbiogenesis

He first notes that biologists have no consensus on life: a paper interviewing 60 experts found 60 definitions. His definition has two halves. The first is von Neumann's: a system that copies itself needs an instruction tape, a machine A that builds what the tape says and a machine B that copies the tape, with the instructions for both A and B on the tape; he says this was worked out in the late 1940s, before the structure of DNA and the discovery of the ribosome and DNA polymerase, and that such a system is a general-purpose computer, so 'biology is computer science'. The second half is symbiogenesis: two or more living things joining into a more complex one that reproduces as a whole.

“life is self-constructing computation that complexifies through symbiogenesis.”Blaise Agüera y Arcas

The other half of evolution is cooperation, not only competition

He says 'random mutation plus survival of the fittest' is half the story and the big steps came from cooperation: mitochondria descend from bacteria taken into cells (shown by Lynn Margulis, whose paper he says was rejected by more than a dozen journals), and the eight 'major evolutionary transitions' listed by Szathmáry and Maynard Smith are all symbiotic. He holds that symbiogenesis is going on all the time: only about 1.5% of the human genome codes for proteins and much of the rest is the remains of retroviruses; mice with Arc, a gene of viral origin, knocked out cannot form new memories; and the placental barrier looks like the product of an ancient viral invasion. So life is more a bush that keeps fusing than a tree, and fitness is not a fixed landscape.

Functionalism: a thing is what it does for other things

He says what separates life from non-life is function: a rock broken in two is two rocks, a kidney broken in two is no longer a kidney. Functions can be carried by other materials: a dialysis machine replacing a kidney passes 'the kidney Turing test', and flight was invented several times in nature. He borrows Chalmers and Schneider's 'brain of Theseus': replace neurons one by one with computers that have the same inputs and outputs, and he holds you would feel no different. So he rejects the idea that a computer cannot be intelligent; efficiency and reliability differ, but not in kind.

BFF: replicators appear on their own in random code

He describes his team's artificial-life experiment: start with very short random programs (he says four bytes) in a self-modifying variant of Brainfuck; put, say, a thousand of them in a soup, take two at random, join them end to end, run them, split them and put them back, again and again. He says that after a few million interactions the programs grow longer, compute a lot and copy themselves; the shortest replicator is about six bytes, longer than one program, so it can only appear by joining (that is, symbiogenesis), and replicators then join into longer, more complex ones. The programs look as if written by 'a mad engineer', and changing one byte at random usually breaks them. He adds that the team has run experiments where programs must chase a light for energy, and machine-learning work where agents that model themselves and others learn to cooperate.

AGI is already here, and intelligence is collective

He says the split between narrow and general AI was made to separate digit recognisers from systems you could talk to about anything, and that large language models, trained on language rather than for a task, have been general intelligence since about 2022; he wrote a piece saying so with Peter Norvig. He says when his team first tested language models on medical diagnosis they were right only about half the time, but doctors given the same descriptions also often erred, and the models did slightly better. He holds that consciousness is theory of mind modelling itself and others in layers, and that human intelligence has always been collective, now including these models.